Re: kill_prior_tuple and index scan costing

Hi,

On 2020-03-21 23:53:05 -0500, Justin Pryzby wrote:
> On Sat, Mar 21, 2020 at 07:33:02PM -0700, Andres Freund wrote:
> > While your recent btree work ensures that we get the heap tids for an
> > equality lookup in heap order (right?),
>
> I think when I tested the TID tiebreaker patch, it didn't help for our case,
> which is for inequality: (timestamptz >= start AND timestamptz < end).
>
> That seems to explain why, although I don't understand why it wouldn't also
> apply to inequality comparison ?

Because tids are only ordered for a single lookup key. So if you scan
across multiple values you could have key:page visited in the order of
1:1 1:2 1:99 2:1 2:7 99:1 or such, i.e. the heap pages would not be
monotonically increasing. You can't however have 1:17 1:1, because for a
specific key value, the tid is used as an additional comparison value.

Greetings,

Andres Freund